978-1-61284-798-6/11/$26.00 ©2011 IEEE 2B5-1

IMPROVING DEPARTURE TAXI TIME PREDICTIONS USING ASDE-X SURVEILLANCE DATA

Amal Srivastava, The MITRE Corporation, Center for Advanced Aviation System Development, McLean, VA, USA

Abstract

Flights incur a large percentage of delay on the ground during the departure process; however, predicting the taxi-out time is difficult due to uncertainties associated with the factors influencing it, such as airport surface traffic, downstream traffic restrictions, runway configuration, weather, and human causes. Airport Surface Detection Equipment, Model X (ASDE-X) surveillance data provides high resolution coverage of aircraft surface movement which can be leveraged to address this problem. This paper presents a novel approach which builds an adaptive taxi-out prediction model based on a historical traffic flow database generated using the ASDE-X data. The model correlates taxi-out time and taxi-out delay to a set of explanatory variables such as aircraft queue position, distance to the runway, arrival rates, departure rates and weather. Two prediction models are developed. One treats aircraft movement from starting location to the runway threshold uniformly while the other models aircraft time to get to the runway queue different from the wait time experienced by the aircraft while in the runway queue. The models are evaluated using data from New York’s John F Kennedy (JFK) airport during the summer of 2010. Results show significant improvement in taxi-out predictions as compared to predictions from FAA’s Enhanced Traffic Management System (ETMS).

Introduction Traffic flow management (TFM) is focused on

balancing demand and capacity to maintain safety while minimizing flight delays. Managing that balance is predicated on predicting both demand and capacity. The greatest source of error in demand prediction comes from error in predicted departure taxi (i.e. taxi-out) time which is defined as the time from the actual pushback to takeoff. Improved departure taxi time prediction will result in more stable trajectory modeling and better en-route sector demand forecasts. It will also help airlines

proactively manage pushback times and improve the efficiency of airport surface movement operations.

Taxi-out prediction for a flight is complex due to uncertainties associated with the factors contributing to it. Surface departure traffic, interaction with arriving traffic, runway configuration, downstream traffic restrictions, airport layout, human causes and weather related delays are all potential contributing factors. Real-time prediction for applications such as trajectory modeling is even more challenging, as data values associated with explanatory variables such as queue position are not readily available. Prior work in the area of taxi-out prediction has been limited due to lack of high quality surface movement data. Predictions have been based primarily on pushback and wheels-off data from sources like the Aviation System Performance Metrics (ASPM) system, and have been made on aggregate quantities rather than the individual flight level. The ASDE-X surveillance system fills this gap by providing high quality surface movement data on aircraft on and around the airport.

This study explores using high-resolution position updates from the ASDE-X surveillance system to develop a taxi-out prediction model. It is based on the premise that existing surface traffic conditions and short-term traffic trends (past 15-60 minutes) revealed from analyzing ASDE-X data are better predictors of future taxi-out time than historical averages. The airport surface is digitized into a number of logical areas such as terminals, taxiways, queues, and runways, and ASDE-X traffic flow data is superimposed on this layout to heuristically determine the aircraft’s location, status (such as taxiing, queued, taking off, airborne, arriving and landed), runway queue order (number of aircraft which will take off before the flight under consideration) and paths of flights on the airport surface. A historical traffic flow database is created and analyzed to discover correlations of potential explanatory variables such as queue position, runway distance, arrival rates, departure rates and weather to the observed taxi-out time. Two approaches to taxi-out predictions are explored. One treats aircraft

2B5-2

movement from the starting location to the runway threshold uniformly (referred to as the “Uniform Flow Model” or UFM) and correlates taxi-out time to a combination of explanatory variables. The other approach, referred to as the “Split Flow Model” (SFM), treats aircraft movement prior to the runway queue (nominal taxi time) differently from when it is in the departure queue (taxi out delay) and predicts taxi-out time as sum of predicted nominal taxi time and predicted taxi-out delay. The models are calibrated for the departure runways of JFK airport and applied to the test data. Predictions from both models are compared to each other and to those obtained from ETMS and historical taxi-out average for a flight. Results show significant improvement in taxi-out prediction compared to the Enhanced Traffic Management System (ETMS) predicted and average taxi-out values. The Split Flow Model (SFM) performs better than Uniform Flow Model (UFM) in traffic conditions when departure queues are present; however, UFM performs more consistently in all traffic conditions.

The organization of this paper is as follows: related work in the area of prediction of departure taxi times is first explored in the Background section. Next section discusses causal factors affecting departure taxi time and how they are included in the study. Next is a high level description of the methodology adopted. Section on Data Sources details the data sources, attributes and data validation strategy used. Following this, the section on Traffic Situation Discovery details heuristics underpinning the determination of flight information. After that is a section on the prediction models and underlying statistical analysis. The results of applying the models to the test data are presented the next section along with analysis and comparisons to ETMS predictions and average taxi-out values. Finally, the conclusion points out the strengths and weakness of this approach.

Background A number of prior efforts have been made to

address prediction of departure taxi times. Shumsky [1] used a linear regression model to predict taxi out time using airline, departure demand and departure runway as explanatory variables. He also developed a queuing model for the runway service process; however the model was based on cumulative

behavior and did not account for the stochastic nature of the process. Herbert [2] modeled the queuing service using exponential, Erlang and Server Absence models, assuming a Markovian system entry process. In another notable work, Idris et el [3] developed a queuing model to predict taxi out time based on factors such as queue size, runway configuration, weather, aircraft type, downstream traffic restrictions and starting terminal. In this study, queue size was estimated based on number of departing aircraft on the airport and estimated number of aircrafts it will pass on its way to the runway. His work concluded that the takeoff queue length is most significant explanatory variable to predict the taxi out time. All these papers identified key factors influencing taxi-out time; however predictions were made at an aggregate level and not on an individual flight basis.

A study on individual flight prediction was done by Pujet [4] who developed a queuing model by predicting the travel time an aircraft will take to reach departure runway and adding that to a runway queue time modeled as a probabilistic service process. Simaiakis and Balakrishnan [5] improved the prediction model by including taxiway interactions as a factor influencing the taxi-out time. In a study based on ASDE-X surveillance data, Legge and Levy [6] modeled taxi-out time based on the relationship with Virtual Departure Queue (VDQ), defined as the number of aircraft which have pushed back before the current one. This study, while using ASDE-X data, did not leverage the detailed traffic situational map that can be constructed using this data. Also, the study limited its scope to using VDQ as the only explanatory variable to taxi-out time.

The approach presented here combines the insights gained from earlier work on causal factors for taxi-out delays with detailed surface traffic situation provided using ASDE-X data to develop a new flight-specific taxi-out prediction model.

Causal Factors and Their Inclusion Consider the airport surface traffic operations

depicted schematically in Figure 1. There is interaction among traffic on different components of the system such as arrival runways, taxiways, terminal gates, ramps and departure runways. As throughput of each component varies, queues are formed in them which add to the taxi-out times. The snapshot of actual airport traffic (JFK on July 9, 2010

2B5-3

at 7:00pm) shown in Figure 2 reveals complex interactions in play among these components. There is a significant aircraft queue waiting to takeoff from runway 22R. Cross traffic from arrivals on runway 22L and departures from runway 13R impact 22R’s departure rates. Taxiing streams of aircraft interact with each other as they merge to enter the runway queue area, causing delays and uncertainty in the aircraft order in the runway queue. Aircraft travel varying distances to the runway, contributing to taxi-out variations. The runway configuration in use alters the traffic pattern and arrival/departure balance and therefore impacts taxi-out times. Downstream traffic restrictions due to weather or high traffic volumes significantly affect departure rates and hence the departure times.

Figure 1. Airport Surface Traffic Operations [7]

Figure 2. Airport Traffic Snap Shot (JFK 7th July 2010 at 7:00pm)

Based on these observations, the following key causal factors are included in this study.

Runway Queue Position It is defined as the number of aircraft which will

depart from the assigned runway before the aircraft under consideration. An aircraft with a zero queue position implies that it will depart next. An algorithm to determine the queue position is described in Traffic Situation Discovery section.

Runway Configuration The departing runway has an impact on taxi-out

times because of variance in taxiing distance and interaction with arriving and departing traffic from other runways. This study models predictions on a runway basis, using it as a primary differentiating factor. The arrival and departure rates of parallel and crossing runways are also included in the analysis.

Airline/Terminal The distance travelled by an aircraft to the

runway varies depending on its starting location, which can potentially impact the taxi-out time. For this study, the straight line distance from the aircraft to the runway threshold is used as approximation of the actual distance.

Weather The impact of weather on taxi-out is included by

using data from Severe Weather Avoidance Programs (SWAP) in the model.

Taxiing Traffic Interactions This is a difficult factor to account for, as there

are very complex traffic interactions during busy times. Runway crossing traffic is accounted for by considering arrival and departure rates from crossing and parallel runways; the arrival rate of the current runway is included to model interactions resulting when the same runway is used for both arrivals and departures.

Downstream Traffic Restrictions Moving average taxi-out time for the runway is

used as the primary indicator of this factor in this study. It represents the traffic flow rate from the runway and is considered an important predictor of

2B5-4

future flow rate. Fifteen, thirty and sixty minute averages are kept.

Methodology Raw ASDE-X data is transformed and enriched in

a series of steps to derive key metrics and generate a traffic flow database which underpin the taxi-out prediction model. The main steps in this process are:

a. Produce a digitized surface layout map of the airport. In this step, airport surface layout features such as runways and taxiways are digitized at sufficient accuracy and stored in XML data files.

b. Traffic situation discovery. The ASDE-X data is cleansed and superimposed on the digitized airport map to infer aircraft status, location, runway used and queue position. A database of historical traffic patterns with attributes representing explanatory variables is created in this step. A traffic visualization framework is developed for display and validation purposes.

c. Validate data. In this step ASDE-X data is validated for completeness and accuracy using visual checks and comparisons with other data sources such as Aviation System Performance Metrics (ASPM) and Airline Service Quality Performance System (ASQP).

d. Prediction model development and application. Linear regression analysis is used to model taxi-out relationship to the explanatory variables. Two different models, namely Uniform Flow Model (UFM) and Split Flow Model (SFM) are developed based on regression analysis and applied to the test data.

e. Results of the study are analyzed to compare performance of the two models and compare them to other sources of prediction such as ETMS and using average taxi-out times for a flight.

Data Sources

Airport Surface Detection Equipment – Model X (ASDE-X) Data

The ASDE-X surveillance system is primarily a runway safety tool that fuses data from multiple sources such as radars on air traffic control tower or remote tower, multilateration sensors, ADS-B (Automatic Dependent Surveillance- Broadcast) sensors, terminal automation system and aircraft transponder into a single stream. The data covers moving “tracks’ which may be aircraft or non-aircraft objects (such as luggage carts) on and around (up to a minimum of five miles from the airport) the airport. There are over hundred data attributes stored for each track. For this study, only the following attributes are used:

� timestamp of the track at one second resolution

� call sign � latitude, longitude, altitude � aircraft type � aircraft heading � speed For this study JFK ASDE-X data from July 3,

2010 to August 14, 2010 is used.

Data Validation The ASDE-X data is evaluated for positional

accuracy, reliability, coverage and completeness. Positional accuracy of data is assessed visually by observing how well tracks of aircraft line up with center-line of taxiways and runways and observing inter-aircraft distances. Figure 3 shows that tracks of two aircraft line up well with the center of the taxiway. Reliability of data is very high as no missing updates were observed in data processing.

Coverage analysis shows that most tracks (95%) start from the taxiway area, just outside the ramp (the “spot”). This implies that the actual pushback event is missed in the available ASDE-X data. Figure 4 shows that ASDE-X spotted the aircraft on average 10.9 minutes after the actual pushback for aircraft headed to runway 22R.

2B5-5

Figure 3. Aircraft Tracks

Figure 4. ASDE-X Start Time Variance with Actual Pushback

The completeness of data was checked by comparing hourly departure figures computed using ASDE-X data with those from ASPM. For JFK during July 2010, ASDE-X reported 94% of flights reported in ASPM. Figure 5 compares ASDE-X and ASPM hourly departures, showing a close match in trends.

Figure 5. Hourly Departure Counts ASDE-X versus ASPM Data

Accuracy of ASDE-X-derived wheels-off time is assessed based on its comparison with Airline Service Quality Performance (ASQP) data. Figure 6 shows the two matched well, confirming accuracy of ASDE-X data.

Figure 6. ASDE-X Wheels Off Variance from ASQP Values (in Minutes)

Overall, ASDE-X data is found to have sufficient accuracy to form the basis of a taxi-out prediction algorithm

Airport Layout &Metadata To trace an aircraft’s surface path, the airport

layout is segregated into logical areas. Google Earth is used to digitize the airport layout information as it is sufficiently accurate and provides flexibility to create named polygon areas which can be saved as files (KML format). The accuracy of this tool is established by visually validating the projected ASDE-X traffic on JFK, EWR, IAD, CLT and ORD airports. Empirical traffic flow patterns are studied to determine the number and layout of “airport areas” needed to capture the traffic flow at adequate

2B5-6

granularity. The following area types are identified and digitized: runways, terminals, taxiways and queues. Queues are the areas where aircraft tend to

line up and wait before taking off from a runway, other areas correspond to their physical equivalents.

Figure 7 shows the airport areas of JFK airport.

Figure 7. JFK Map with Digitized Airport Areas

Apart from the layout coordinates of the areas, additional information is collected to aid processing. Each area has one or more adjoining areas; centers of the dividing edges are captured as exit-points for that area. Figure 8 shows four exit points from Taxiway 7. Depending on the number of exit-points, an area may contain multiple flows of aircraft.

Figure 8. Exit Points for Taxiway 7

2B5-7

Note that there can be multiple queues for a runway, for example Queue1 22R and Queue2 22R serve runway 22R (see Figure 7). The layout information along with additional metadata is captured in an airport-specific configuration file

Traffic Situation Discovery

Flight Status and Location Positional attributes of ASDE-X data are

mapped to areas configured in the digitized airport layout to locate it on or around the airport. Aircraft spotted first while on the airport surface are considered to be departing, while those in the air (altitude above ground level is greater than zero) are considered arriving. An aircraft is marked as being queued when it is in the queue area and its speed drops below a threshold. A flight is marked as taking off when it is on the runway and its speed exceeds a takeoff speed threshold; once its altitude exceeds zero it is marked as having taken off and is airborne. Similarly, an arriving flight is marked as landed when its altitude drops to zero. Aircraft location and status feed into the process of queue assignment and ordering.

Queue Assignment and Order Each departing flight is assigned to one of the

currently active runways. If there is only one active departure runway, all departing flights are assigned to its queue. In the case of multiple active runways, the algorithm assigns an aircraft to the runway X queue as follows:

� aircraft on the runway X are assigned first � aircraft on the queue area to runway X are

assigned next � aircraft on area adjacent to the queue area

are assigned to runway X next. � aircraft heading to the runway X queue are

assigned next depending on the number of airport areas they have to traverse to get to the queue area.

In Figure 9, white aircraft icons are assigned to runway 22R and black to runway 13R. Note that an aircraft’s proximity and heading are both considered for queue assignment; for example, the aircraft at

queue position three is assigned to runway 13R even though it is close to the queue for 22R.

Figure 9. Queue Ordering for 22R & 13R

Queue order is determined as follows:

� aircraft on the runway are ordered based on their distance to the departing runway end.

� aircraft on queue are ordered next, based on their distance to runway entrance point

� aircraft on other surfaces are ordered next based on their distance to runway entrance point

Runway Activity Data Runway activity is recorded every minute to

compute arrival rate, departure rate and average taxi-out time over the last fifteen, thirty and sixty minutes.

Traffic situation discovery results are captured in a database table. Key attributes collected are: call sign, starting area, start time, surface path, taxi-out time, departing runway, initial runway assignment, initial queue position and initial distance to the runway. For SFM, queue entry time, queue position and distance to runway on queue entry are also captured.

Taxi-Out Prediction Models The taxi-out prediction model is based on

modeling the relationship of taxi-out time to the explanatory variables described previously. Historical traffic data from 3 July 2010 to 23 July 2010 is

2B5-8

processed to create the database used to calibrate the models.

Figure 10 shows a typical aircraft taxi-out track. Using traffic situation discovery rules, the following data points are collected for each flight:

� time when the aircraft is first spotted � runway assignment � queue time when the aircraft stops in

queue area indicating it is waiting in queue � wheels off time � initial queue position in the runway queue � queue position in the runway queue when

the aircraft gets queued � straight line distance of aircraft to the

runway threshold

Runway 22R

Queue1 22R

Taxiway 4B

Taxiway 4A-5

Taxiway 5

Start Time

Queue Time

Take Off

Runway Distance

Figure 10. Data Points Collected for Each Flight

The following runway activity data is collected every minute:

� active arrival and departure runways � moving taxi-out average for each departure

runway in fifteen, thirty and sixty minute bins

� arrival and departure rates for each runway in fifteen, thirty and sixty minute bins

Taxi-out time is modeled using two approaches.

Uniform Flow Model (UFM)

This approach assumes uniform aircraft movement behavior during the entire taxi-out process

and models the taxi-out time in a single segment.

Taxi-out time is modeled as function of explanatory variables as discovered using linear regression analysis. The model can be applied to all flight departures.

Split Flow Model (SFM) This approach assumes that aircraft movement

behavior when it is travelling to the queue is different from when it is waiting in the queue. The model calculates taxi-out time as sum of time to get to the queue (nominal time) and time in the queue.

Taxi-Out Time = Nominal Time + Queue Time A nominal taxi-out time value for each runway

is established as the ten percentile point of the actual taxi-out time distribution. Figure 11 shows the distribution for runway 22R, 6.43 minutes is used as nominal time for this runway.

Figure 11. Actual Taxi-Out Distribution for 22R

Exception Situations At times, an aircraft takes inordinately large

taxi-out time due to exception conditions such as mechanical issues or human causes. Taxi-out values ������ ���� � ��� ����� � � � ��� �� � ����� �� ��account for these situations. In some cases the taxi-out predictions can be impacted due to anomalies in traffic pattern, such as aircrafts cutting through the queue. To account for that, five percent of outlying predictions are ignored when considering the performance of all prediction methods.

2B5-9

Model Equation The regression analysis on historical database

fits the data to an optimal linear solution, which outputs predicted taxi-out time (taxi-out delay for SFM) as a function of the explanatory variables. This is represented as a linear equation for each departure runway considered. For UFM, the equation produces the final predicted taxi-out time; for example, equation for the runway 4L is:

Ttaxiout = 151 + 64.1*QP + 0.10*RD + (1)

0.03*FA + 38.05*4LA + 13.64*4RA

Ttaxiout Predicted taxi-out time (seconds)

QP Queue position

RD Distance to runway start (meters)

FA Taxi-out mean (15 minutes bin)

4LA 4L Arrivals last fifteen minutes

4RA 4R Arrivals last fifteen minutes

For SFM, taxi-out time is computed as sum of queue time (or taxi-out delay) and nominal time to get to the queue. The queue time is predicted using linear equation derived from regression analysis, such as equation (1) and nominal time is taken as ten percentile value of taxi-out distribution for the runway.

Model Prediction Results & Analysis UFM and SFM models were applied to JFK

ASDE-X data from 24th July 2010 to 13th August 2010 to assess the following:

� Impact of explanatory variables on taxi-out time

� Comparison between the two ASDE-X prediction models

� Stability of ASDE-X model predictions � Comparison between ASDE-X and ETMS

predictions and average taxi-out values Table 1 shows mean, standard deviation and

coefficient of determination (R2 or square of correlation coefficient) of taxi-out prediction error (predicted-actual taxi-out) in minutes for the two models broken down by explanatory variable for each

departure runway. Variables taxi-out mean and all arrival rates are for fifteen minute bins; thirty and sixty minute bins are ignored as they are less significant. The last row of each set (All) is based on a multi-linear regression of all the explanatory variables in the set.

Table 1. Runway Wise Model Performance

Runway 4L

Variables Uniform Flow Model Split Flow Model

Mean Stdev R2 Mean Stdev R2 Queue Position 0.31 3.91 0.4768 0.50 3.61 0.7361

Taxi-out Mean 0.05 5.19 0.1516 0.28 5.18 0.1463

Runway Distance 0.59 5.12 0.0464 0.65 4.78 0.4055

Severe Weather n/a n/a n/a 0.67 5.38 0.0032

4L Arrivalrate 0.68 5.05 0.0333 0.80 5.21 0.0199

4R Arrivalrate 0.68 5.37 0.1550 0.80 5.30 0.1697

All 0.37 3.53 0.5423 0.50 3.51 0.7903 N=1525 Nominal Time=7.57 min

N=2994 Nominal Time=9.83 min

Runway 13R

Variables Uniform Flow Model

Split Flow Model

Mean Stdev R2 Mean Stdev R2 Queue Position

-0.32 5.79 0.5173 1.20 6.63 0.7216

Taxi-out Mean

-1.75 7.18 0.2338 -0.27 7.83 0.2839

Runway Distance

-4.68 9.50 0.0065 -2.07 9.0 0.3379

Severe Weather

-3.7 9.05 0.0205 -1.4 9.4 0.0013

13R Arrival rate

-4.67 9.46 0.0066 -1.49 9.5 0.0348

13L Arrival rate

4.72 9.54 0.0025 -2.45 9.78 0.0311

All -0.40 5.82 0.5280 0.10 7.35 0.7775

2B5-10

Runway 22R

Variables Uniform Flow Model

Split Flow Model

Mean Stdev R2 Mean Stdev R2 Queue Position

1.77 5.11 0.5671 0.51 5.56 0.6918

Taxi-out Mean

1.56 6.11 0.4945 0.67 7.28 0.1487

Runway Distance

4.24 7.64 0.0407 0.54 6.23 0.3183

Severe Weather

3.50 7.86 0.1486 1.13 7.85 0.0654

22R Arrival rate

4.31 8.22 0.0146 0.90 8.20 0.0288

22L Arrival rate

3.92 7.82 0.0314 1.20 7.91 0.0441

All 1.15 4.89 0.6447 0.51 5.55 0.7209 N=2206 Nominal Time=6.43 min

N=1685 Nominal time=5.83 min

Impact of Explanatory Variables Queue position emerges as the most significant

determinant of taxi-out time as indicated by highest value of R2 for both the models. For UFM, average taxi-out time is next, followed by arrival rates on parallel runway. For SFM, runway distance is the next most significant factor, followed by average taxi-out time; other factors are insignificant.

Figure 12 shows a scatter plot of UFM predicted taxi-out times against queue positions for runway 4L; linear and quadratic fits are plotted. The two plots are almost identical, indicating that a linear relationship can be assumed for this most significant explanatory variable. Figures 13 and 14 show the scatter plot of average taxi-out time (15 minutes) and 4R arrival rates versus the taxi-out times, indicating a less well-defined relationship compared to queue position. Arrivals on parallel runways are more significant for runways 4L and 31L than other runways. Since the models consider already taxiing flights and make short-term predictions, weather did not emerge as a significant factor in this study.

Figure 12. Scatter Plot of Taxi-Out Time and Queue Position with Linear and Quadratic Fit

Curves

Figure 13. Scatter Plot of Taxi-Out Time and 4R Arrival Rates (15 Minute Bin)

Runway 31L

Variables Uniform Flow model Split Flow Model Mean Stdev R2 Mean Stdev R2

Queue Position

0.83 6.53 0.5356 -7.11 10.80 0.1621

Taxi-out Mean

-0.10 7.33 0.3504 -7.12 10.83 0.0237

Runway Distance

0.20 8.96 0.1995 -7.12 10.88 0.1066

Severe Weather

-0.12 9.25 0.0024 -7.11 10.83 0.0096

31L Arrival rate

0.84 8.51 0.2295 -7.11 10.85 0.0003

31R Arrival rate

1.18 8.75 0.1554 -7.12 10.85 0.0133

All 0.83 6.05 0.6176 -7.11 10.83 0.3540

2B5-11

Figure 14. Scatter Plot of Taxi-Out Time and Taxi-Out Average Last 15 Minutes

Models Comparison The Split Flow Model provides better prediction than Uniform Flow Model for runways 4L, 22R and 13R as indicated by a higher coefficient of determination. This is supported by empirical observations on aircraft movement behavior when they are queued during departure. For runway 31L, the Split Flow Model performs significantly worse as shown by lower coefficient of determination and higher spread and mean error values. This is due to significantly less time the flights were queued for runway 31L (~15 percent) compared to other runways (~50 percent) as shown in Table 2. UFM performs more consistently than SFM; however SFM performs better in traffic conditions where queues develop during the departure process.

Table 2. Relative Queue Times for Runways Runway Taxi-out time Queue time

Mean Std Dev Mean Std Dev 4L 15.38 8.17 8.61 5.38 22R 22.11 13.31 10.95 9.39 13R 20.34 10.77 11.34 7.91 31L 16.39 16.38 2.38 2.29

Model Stability Model stability is assessed by plotting the

impact of varying traffic conditions on prediction quality, measured as the absolute difference between predicted and actual taxi-out times as percentage of taxi-out time. Average number of push backs and average taxi-out time are used as measure of departure traffic load and delay. Figure 15 shows traffic variation and taxi-out error variance percent

slotted on identical times of the day; all the values are averaged in thirty minute bins. UFM model predictions on JFK data from 24th July 2010 to 13th August 2010 are used for this analysis. The charts show that predictions remain fairly stable in varying traffic conditions.

0

5

10

15

20

25

30

0:00

1:00

2:00

3:00

4:00

5:00

6:00

7:00

8:00

9:00

10:0

011

:00

12:0

013

:00

14:0

015

:00

16:0

017

:00

18:0

019

:00

20:0

021

:00

22:0

023

:00

Traffic Variation

Pushbacks Taxiout

0

20

40

60

80

100

0:00

1:00

2:00

3:00

4:00

5:00

6:00

7:00

8:00

9:00

10:0

011

:00

12:0

013

:00

14:0

015

:00

16:0

017

:00

18:0

019

:00

20:0

021

:00

22:0

023

:00

Taxiout Error Variance Percent

Taxiout error variance

Figure 15. Average Daily Traffic Trend versus Taxi-Out Error Variance

Model Comparison with ETMS and Averages ETMS generates a series of predictions for a

flight starting from the time it is filed to when it takes off. For this analysis, UFM predictions are compared to the ETMS predictions active at that time. The runway-wise taxi-out error in minutes (predicted minus actual taxi-out time) distribution is depicted in Figure 16. The ASDE-X based model significantly improves predictions for all runways for both taxi-out error mean value and reduced error spread (����� ����In another comparison, taxi-out prediction for a flight is calculated as average of ten or more taxi-times for the same flight during the period from 3 July 2010 to 23 July 2010 (Averages model). Last chart of Figure 16 shows taxi-out error distributions for ETMS, Averages and ASDE-X model for all the runways. The ASDE-X based model significantly improved predictions over other models both in reduced mean ����� ���� � �� �� ����� ��� ��� ��� � ��� �� ��� ��values.

2B5-12

Figure 16. Runway Wise and All Runways Comparison between ASDE-X, ETMS and Averages Models

Runway 4L ETMS Predictions ASDE-X Model Predictions

-20 -10 0 10 20

MeanStd DevStd Err MeanUpper 95% MeanLower 95% MeanN

0.64767295.42022670.11304410.86935210.4259937

2299

Moments

Runway 22R ETMS Predictions ASDE-X Model Predictions

Runway 13R ETMS Predictions ASDE-X Model Predictions

Runway 31L ETMS Predictions ASDE-X Model Predictions

All ETMS Predictions Averages Predictions ASDE-X Model

2B5-13

Conclusion This paper introduces two taxi-out prediction

models based on high resolution airport surface movement data available from ASDE-X surveillance system. Uniform Flow Model (UFM) assumes uniform aircraft movement from the starting location to the runway threshold and models taxi-out time in a single segment. Split Flow Model (SFM), on the other hand, models taxi-out time as sum of predicted time to reach the runway queue and taxi-out delay while the aircraft is in the queue. Linear regression analysis on traffic flow database generated using ASDE-X data is used to model taxi-out time (in case of UFM) or taxi-out delay (in case of SFM) as a variable dependent on explanatory variables such as queue position, runway distance, arrival rates, departure rates and weather. Results from applying the models show a significant improvement in quality of predictions when compared with predictions currently available from ETMS or using average taxi-out times (Averages model) for a flight. ASDE-X model (UFM) reduced mean prediction error (predicted-actual taxi-out time) from -8.07 (ETMS) and 2.29 (Averages) to 0.68 minutes. Error spread (standard deviation) was reduced from 21.78 (ETMS) and 8.13 (Averages) to 4.61 minutes. Comparison between the two models shows that SFM performs better in traffic conditions when queues develop at runways, while UFM performs more consistently in all traffic conditions. This study also shows that queue position is most significant determinant of taxi-out times.

Since ASDE-X updates traffic data every second, the model can update predictions as aircraft approach runway, making it suitable for deployment in a real-time environment.

One weakness of this model is that the predictions are for a short-term horizon (thirty minutes or so into the future), as they can only be made when the aircrafts come into the ASDE-X view. The ASDE-X data used in the study starts tracking flights when they leave the ramp area. If the ASDE-X data within the ramp area were available, modeling can start at pushback, thus extending the look ahead time. Also, the model needs to be recalibrated if traffic pattern changes significantly, such as due to closing of a runway, or during winter when aircraft wait for de-icing. The model should

constantly adapt to changing traffic pattern by monitoring the error rates and recalibrating if needed.

This model can be further improved by considering other factors such as departure separation rules and fleet mix, flight’s departure clearance time when an Expected Departure Clearance Time (EDCT) has been assigned, and the first departure fix in the flight’s filed route to fine tune the queue ordering and wait times.

References [1] Shumsky, R.A. [1995], Dynamic Statistical models for the prediction of aircraft take-off times, Ph.D. Thesis Operations Research Center, MIT, Cambridge, MA.

[2] Herbert, E.J. and Dietz, D.C. [1997] Modeling and Analysis of an Airport Departure Process,Journal of Aircraft, Vol. 34(1) January-February 1997.

[3] Idris, Husni R., John-Paul Clarke, Rani Bhuva,and Laura Kang, September 2001, Queuing Model for Taxi-Out Time Estimation, ATC Quarterly.

[4] Pujet, N., Delcaire, B. and Feron, E. [1999], Input-Output Modeling and Control of theDeparture Process of Busy Airports, MIT International Center for Air Transportation, #ICAT-99-4

[5] Simaiakis, I. and Balakrishnan, H., Queueing Models of Airport Departure Processes for Emissions Reduction. In AIAA Guidance, Navigation and Control Conference and Exhibitm 2009

[6] Legge, J. and Levy B., Departure Taxi Time Predictions Using ASDE-X Surveillance Data, 26th International Congress of Aeronautical Sciences 2008

[7] Idris, Husni R., Observations and analysis of departure operations at Boston Logan International Airport. PhD thesis, Massachussetts Institute of Technology, 2001

Acknowledgments The author is grateful to Dr Craig Wanke

(Senior Principal Engineer) and Timothy Stewart (Principal Engineer) at MITRE Corporation for their valuable suggestions and feedback on this paper.

2B5-14

Disclaimer The contents of this material reflect the views of

the author and The MITRE Corporation and do not necessarily reflect the views of the FAA or the DOT. Neither the Federal Aviation Administration nor the Department of Transportation makes any warranty or

guarantee, or promise, expressed or implied, concerning the content or accuracy of these views.

30th Digital Avionics Systems Conference October 16-20, 2011